ECN publication
Agent-oriented modeling. An initial investigation.
Published by: Publication date:
ECN Policy Studies 3-1-1994
ECN report number: Document type:
ECN-I--94-022 ECN publication
Number of pages: Full text:
74 Download PDF  

In this thesis an account is given of an initial investigation of a certain class of integrated models, called agent-oriented models. Integrated models are defined as models that consist of several submodels, either linked by the exchange of values of parameters, or embedded in one large model. Agentoriented models belong to the first group. The research has been carried out for Energy Research Foundation ECN. The business unit ECN Policy Studies takes an interest in integrated energy models, because of potential uses of such models for description of EG markets, description of the interaction between the energy sector and the rest of the economy, and combinations of technologically oriented models with behavioral models. From a literature survey it was concluded that existing integrated models suffer from a lack of transparency. This motivated the investigation of properties of a class of integrated models characterized by a modular design. Each submodel describes the optimizing or simulating behaviour of only one agent. Agents take the state of their economic environment as given. A few examples of agent-oriented models, based on microeconomic situations, give an illustration of the principles. Conditions for the existence of equilibrium (the state in which all agents have consistent optimal solutions) and the properties of iterative solution methods for agent-oriented models are investigated. For this purpose, the models ere viewed as systems of nonlinear equations which represent the response functions. It is shown that an equilibrium can be interpreted as a fixed point to this system of response functions. The conditions for the existence of equilibrium are quite strict. In order to prove convergence of iterative methods, even stricter claims must be made on the response functions. As an alternative approach, the use of a piecewise linear approximation of the response functions for all component models is proposed. Iterative methods can be used to solve the system of approximated response functions. For one of these methods it is proved that convergence to a fixed point always occurs, provided that a fixed point exists in the integrated model consisting of approximated response functions. Some experiments were carried out in order to get some practical experience with this modeling methodology. An adapted version of an model of the European gas market was implemented in GAMS as an agentoriented model. Piecewise linear approximations were incorporated for the models of gas producing countries. From the experiments, it became clear that the methodology is not so flexible that any model with appropriate inputs and outputs can just be plugged in, and an equilibrium can be found. On the contrary, agent-oriented modeling only makes sense if the response functions are defined in such a way that the inputs have some control over the outputs. A computational disadvantage was formed by the execution time needed to solve the GAMS version of an integrated model. It is concluded that agent-oriented modeling should only be used if in the real life situation all variables that serve as inputs into different component models have a sufficient influence on the variables that serve as outputs from these models. Then agent-oriented modeling yields many advantages, like transparency, modularity, and, when approximation is used, the possibility to incorporate large component models without having to solve them in each iteration. The use of approximations of the response functions extends the usefulness of the approach. For large agent-oriented models, the applicability of the methodology would be considerably improved by the development of a modeling methodology which takes care of file handling and approximations.

Back to List